Memory management for multi-application managed runtime environments

Modern computing platforms are pervasive, networked, heterogeneous and increasingly complex. These systems range from small hand-held devices such as cell phones, to large servers that run back-end software. Portability, widespread use, and security of program execution as well as support for a fast program development cycle are key concerns of software developers and end users of these platforms. To address these concerns, modern programming languages and their implementations have emerged to facilitate high-level, object-oriented, and type-safe software development that is portable and secure. Popular examples of these languages are Java and the .NET languages. Programs written in these languages are encoded by a source compiler into an architecture-independent format, and then executed on any platform for which there is a virtual execution environment called a Managed Runtime Environment (MRE). MREs provide dynamic compilation of programs to the underlying native machine format, adaptive optimization, memory management, security verification, and other runtime services for programs. Automatic memory management, or Garbage Collection (GC), is an MRE service that is key to facilitating programmer productivity, portability, and memory safety of programs written in Java and the .NET languages. GC relieves the user from employing (and debugging) explicit deallocation of heap memory. However, since such management and reclamation is provided by the MRE, it necessarily introduces overhead. Much prior work has focused on mitigating the overhead of GC for MREs that execute a single program at a time using a single operating system process. However, today's MRE platforms execute a wide variety of applications with diverse computational characteristics, resource requirements and object lifetimes. Moreover, state-of-the-art MREs now are able to execute multiple applications using a single MRE process, serially or concurrently. For widespread use of such systems, and arguably for the success of such languages, we must advance GC technology to exploit the multi-program execution model and to enable efficient automatic memory management and program performance for the next-generation of portable software. In this dissertation, we focus on memory management techniques for the next-generation of multi-program MREs for Java and consider both serial (persistent) and concurrent program execution models. In particular, we investigate, design, and engineer two novel solutions for GC that facilitate high-performance program execution for these MREs: (i) Application-Specific Garbage Collection and (ii) Scalable Memory Management for Multi-Tasking MREs. In the former, we customize and automatically specialize the GC in an MRE for a particular application and switch GCs automatically within a single, persistent MRE as needed for each new application that the MRE executes. In the latter, we present a novel GC system for concurrent execution of programs; this system isolates GC activities of individual tasks while programs share a single heap. Our results show that we are able to achieve significant performance gains over the state-of-the-art MRE systems for both research and production-quality multi-program MRE systems in terms of throughput, response time, and memory footprint.